Superpower Daily
Google Agrees to Buy 396 MW of Geothermal Power for a Planned U...Anthropic’s Fable 5.1 Takes Benchmark Lead, but Its Top Setting...LlamaIndex and Kaggle Launch a Document Benchmark for Cost and...Meta Rolls Out Muse Spark 1.3 for Longer Tasks, With Max Reason...CrowdStrike Wants AI Agents Rechecked Before Every ActionLoudoun County Confronts Data-Center Backlash After Reaping a T...Google Agrees to Buy 396 MW of Geothermal Power for a Planned U...Anthropic’s Fable 5.1 Takes Benchmark Lead, but Its Top Setting...LlamaIndex and Kaggle Launch a Document Benchmark for Cost and...Meta Rolls Out Muse Spark 1.3 for Longer Tasks, With Max Reason...CrowdStrike Wants AI Agents Rechecked Before Every ActionLoudoun County Confronts Data-Center Backlash After Reaping a T...
Productspublished

CrowdStrike Wants AI Agents Rechecked Before Every Action

The company’s approach makes an agent’s authority temporary: access is scoped to a job and can be withdrawn when that job ends or risk conditions change.

By 2 min read
CrowdStrike Wants AI Agents Rechecked Before Every Action
CrowdStrike Wants AI Agents Rechecked Before Every Action
Sign in to save and react

Listen to this story

The audio brief

About 1:37
0:001:37
Read transcript
CrowdStrike is turning AI-agent access into a sequence of decisions, checked before each action rather than granted for an entire login session. The company highlighted that model, called Continuous Identity for AI Agents, at its Fal.Con 2026 keynote in Las Vegas. Its system gives an agent short-lived authority for a specific task, then can withdraw it when the task ends or the surrounding risk changes. Before approving an action, it considers who owns the agent, who called it, the device’s risk, the user’s entitlements, and the current business context. That means a read-and-write agent acting for someone with read-only access can read, but not authorize a write. CrowdStrike also says delegated sub-agents preserve the initiating human’s identity and permissions. The capability followed CrowdStrike’s seven-hundred-forty-million-dollar acquisition of SGNL, announced in January, and was productized in June. It uses SPIFFE and the Shared Signals Framework to support identity and real-time authorization changes. A second layer, Falcon AI Detection and Response, continuously inspects prompts and intent for misuse or attempts to push a language model beyond its allowed scope, with revocation possible before damage occurs. CrowdStrike says agents are producing roughly 250 times more detections than humans, while average eCrime breakout time has fallen to 29 minutes, with a fastest recorded intrusion of 27 seconds. The key constraint is whether these checks can keep pace as an agent’s context changes mid-task.

Story brief

3 key points

CrowdStrike is positioning Continuous Identity for AI Agents as an authorization layer that can change throughout an agent’s work, not just at login. The capability, built after its $740 million SGNL acquisition and productized in June, evaluates human and agent permissions alongside device and security context, then can revoke access mid-task. CrowdStrike says Falcon AI Detection and Response also monitors prompts...

  1. 01

    CrowdStrike announced its $740 million SGNL acquisition in January and productized the capability as Continuous Identity for AI Agents in June.

  2. 02

    Authorization considers the agent owner, caller, device risk, entitlements, and current business context before each action.

  3. 03

    CrowdStrike says delegated sub-agents retain the initiating human’s identity and permissions; read-only users cannot authorize writes.

An AI agent’s access can now be treated as a series of decisions rather than a privilege that lasts for a full login session. CrowdStrike highlighted that model, called Continuous Identity for AI Agents, during its Fal.Con 2026 keynote program in Las Vegas.

Access that expires with the job

CrowdStrike’s system continuously authorizes agent actions instead of trusting an agent for an entire session. It grants short-lived authorization for specific tasks and can revoke access when a task ends. Before approving an action, it evaluates the agent’s owner, the caller and the device’s risk posture.

A decision can be reversed

The design’s key operational feature is revocation after an initial approval. CrowdStrike says it can immediately remove access when the context changes, including when a vulnerability is discovered or an employee’s HR status changes. The company says the capability uses SPIFFE, an identity standard, and the Shared Signals Framework for real-time authorization.

The company’s stated control model

  • Every action is assessed against both the agent’s and the human’s entitlements, plus current security and business context.
  • A read/write agent acting for a read-only user can only read, CrowdStrike says.
  • When an agent delegates to a sub-agent, the company says the human identity and permissions are preserved.

A second check on agent behavior

CrowdStrike says the capability is delivered through Falcon Next-Gen Identity Security and Falcon AI Detection and Response. The latter continuously inspects prompts and intent for permission misuse or attempts to manipulate a language model beyond its authorized scope, and can trigger revocation before damage is done.

The company’s case for speed

CrowdStrike researcher Adam Meyers said AI agents are generating roughly 250 times more detections than humans. Separately, CrowdStrike threat research cited in the article found that average eCrime breakout time had fallen to 29 minutes, with the fastest recorded intrusion at 27 seconds. Those company figures frame the push for access controls that can change while an agent is working.

Sources

  1. siliconangle.comContinuous AI agent identity becomes a 24/7 job at Fal.Con - SiliconANGLE
  2. crowdstrike.comwww.crowdstrike.com

Related coverage

Continue reading